Autoencoder-FCNN hybrid neural network for optical alignment in real-world applications

Accurate alignment of optical components is critical in large optical systems such as astronomical and aerospace/defense instruments. In such systems, achieving high-quality wavefronts is essential for long-range resolution and beam stability. However, the physical scale introduces practical challenges to alignment, with conventional methods relying on labor-intensive mechanical adjustments and iterative measurements requiring over 20 iterations and being susceptible to environmental noise. To address these limitations, this study introduces a hybrid unsupervised-supervised deep learning model, Autoencoder - Fully Connected Neural Network (AE-FCNN), which directly estimates misalignment parameters from interferometrically measured wavefronts. The AE separates stable errors from measured data, while the FCNN predicts misalignments from the residual wavefront features. The model was trained using simulated wavefronts, including experimentally measured figure errors and synthetic random noise, and validated on real-world interferometric data. Alignment was achieved within 2-3 iterations in all cases, reducing adjustment steps from 6 to 2 in the prototype optical system and from 15 to 2 in the large optical system. This demonstrates that AE-FCNN improves alignment efficiency while maintaining robustness to experimental noise. The proposed model does not require domain-matched datasets, enabling training on simulations and application to experimental data. This highlights its practicality for real-world large system alignment tasks.